Dynamic trading strategy learning model using learning classifier systems

Current trading strategy learning models often proceed in three separate phases, i.e., training, validation, and application (testing). After a specific time span of application, a new learning process is started to adapt the trading strategy to the new environment states. The time span of application is usually fixed and determined according to experiences. This may result in earning losses as compared to the perfect trading strategy which trades at each turning point of the stock price movement. Some learning methods, such as neural networks, are hard to explain intuitively and unstable in some dynamic environment states. Other learning models like simple genetic algorithms result in a single trading rule which is applied for a specific time span without being adapted even when the environment has changed. This paper adopts learning classifier systems (LCSs) technique to provide a dynamic trading strategy learning model (DTSLM), which makes continuous and instant learning while executing real prediction and produces a trading rule set to deal with different environment states. The simulation results show that this model could get a remarkable trading profit.